Improving the Performance of Inconsistent Knowledge Bases via Combined Optimization Method

One of the important issues when designing effective expert systems is the validation and refinement of the acquired knowledge bases. The validation and refinement problem becomes more important and more difficult when the knowledge bases of expert systems consist of uncertain rules, e.g., probabilistic rules. In this paper, we first describe one type of inconsistency of knowledge bases called sociopathicity and summarize some results. We then develop the Combined Optimization Method to debug inconsistent knowledge bases. This method utilizes the static and dynamic information of the rules. Our experiments show that the debugged knowledge bases by the method significantly improve the system performance on the validation sets as well as on the training sets. The experimental results also empirically verify the manifatation of the sociopathic interactions among the rules and the improbability of locally debugging this type of inconsistent knowledge bases. tical admissibility, modification to the knowledge base being conservative (because we are not in the phase of learning or overhauling the entire knowledge base.) In this framework, many researchers have proposed methods to refine knowledge bases in various ways [Ginsberg et al., 19881. A more focused area of the knowledge base refinement problem is to detect the inconsistency and incompleteness of knowledge bases. Some techniques have also been developed under different representations [Ginsberg, 1988, Kim, 1988, Nyugen et al., 1985, Suwa et al., 1982). In particular, the inconsistency checking is very important because inconsistent knowledge bases can and do lead to wrong conclusions which is an undesirable consequence. This paper studies one special type of inconsistency that exists in knowledge bases of uncertain rules their sociopathicity [Wilkins and Buchanan, 19861. It has been shown that the problem of optimizing the performance of sociopathic knowledge bases is NP-hard [Wilkins and Ma, 19911. In this paper, we will develop the Combined Optimization Method to refine inconsistent knowledge bases so as to improve their empirical performance. The experiments show that the method significantly improves the performance of knowledge bases both on the training sets and on the validation sets.